Census the obligation scope gap: ten of twelve grammar positions have no producer

[gunbai-bot]

Continuation of #9879 (merged). Same class, same row — one commit, one file, no code.

This commit was written against #9879's head and stranded when that PR merged: under squash-merge a branch tip is never an ancestor of main, so the merged branch kept accepting pushes. Carried onto a fresh branch from current origin/main.

The finding

v1.compiler.04_infer DeclaredTypePosition declares twelve grammar positions that can receive a source value. DeclaredTypeObligation is constructed at exactly two sites — PositionDirectCallArgument and PositionListElement. Every other member occurs only in declared_type_position_label, which renders a name and judges nothing.

Found by neat-hawk-884; verified here independently of the carrier's own annotation by counting construction sites of that type in that module: one declaration, two constructions.

Why the row points at the carrier instead of restating it

The carrier already states this gap in full §4b shape in its own voice — the enumeration of the ten, the refusal to file them as InhabitanceUndecidable, and a per-position trigger. DESIGN §6 says the mark on the carrier is the authority and a parallel-ledger doc is the failure mode, so copying that content into the census document would mint the second authority the row exists to prevent. Recorded here is only what the carrier does not say:

1. It is a THIRD cause of silence

The board has been reasoning with two — the judgment ran and accepted, or ran and declined. The bodyless-type finding in #9879 established the first is real. This adds no producer at that seam at all. The three are indistinguishable from outside the compiler, which is why an admission cannot be attributed to a cause without reading the wiring — neat-hawk-884 reports 45 of 64 board rows UNDECIDED on exactly this.

2. Occupancy per position — which is NOT the same question as reachability

Nine of the ten are occupied today, each by a cited specimen rather than a count.

PositionParameterDefault is grammar-admitted (v1.compiler.02_parse parses name: Type = expr and is the only path constructing a param node with a non-none default) and corpus-unoccupied: zero parameter defaults across 42,329 fn declarations, paren-balanced scan, detector validated against four planted positives before the zero was trusted — a silently-broken detector returns the same zero.

That zero does not make the position inert. Reachability is judged by what a fixture may author, never by what the accepted corpus contains (DESIGN §4b; the sibling ledger's reachability_read_as_occupancy row). An unoccupied but grammar-admitted position has an authorable RED: no exposure today, exposure the moment someone writes the construct, with nothing to refuse them. So the ten stay ten, split as exposure-now vs exposure-on-first-use rather than real vs inert.

3. Ceilings do not differ by position, and the reason matters more than the answer

None of the ten is grammar-absent, so none is unreachable for that reason. The decision procedure is shared and position-agnostic, so each unwired position is a construction site rather than a rule — which is why the carrier prices the residual ten at a call each.

The one place a ceiling genuinely bites is inside the relation, not at a position: InhabitanceUndecidableReason already carries UndecidableGenericFormal, so PositionGenericTypeArgument reaches a decidability limit the others do not — a property of the facts rather than of the wiring. R2 is not promised there.

Disposition

• Rung below floor for all ten — absent from the census, no diagnostic of either severity.
• Ceiling R2, same as the two wired positions, for the reason above.
• Next trigger at capability grain: every grammar position that can receive a source value constructs a DeclaredTypeObligation and hands it to the shared decision procedure. The carrier's ten per-position triggers are the steps; satisfying one retires a step, not this row.
• NOT ENROLLED by ruling, not omission (bright-ram-778): enrolling an expected red is a decision about the required gate and belongs to the owner of v2.workflow.floor_expected_red, not to the author who found the gap.
• No producers written. This is a census. Where a producer is trivially addable that is evidence for the trigger, not licence to add it.

🤖 Generated with Claude Code

https://claude.ai/code/session_013wJU7jqGAEk8rj4UwpHdox

[gunbai-bot]

review 58065 is right, and the rule it cites disposes of more than the figure it caught. Fixed in e7987ca82f1.

The finding, verified. The row transcribed "42,329 fn declarations" and a zero from a paren-balanced scan I ran once in a session worktree. There is no committed producer and no entry point that re-derives either number. DESIGN §6 is explicit — a measurement is cited by naming the producer that re-derives it, never by copying its output — and its own test settles which side this falls on: occupancy of a grammar position is worth re-deriving, since it changes the day someone authors the construct, so it is worth an entry point and is therefore an instrument the row is owed, not a number the row may carry.

What replaced it is stronger, because the conclusion never depended on the count. Reachability is judged by what a fixture may author, never by what the accepted corpus contains. So a grammar-admitted position has an authorable RED whether or not the corpus occupies it: an empty corpus means no exposure today and exposure on first use, a non-empty one means exposure now, and neither reading makes the position inert or moves the rung, the ceiling, or the trigger. Grammar admission itself is decidable by symbol and stays cited (v1.compiler.02_parse parses name: Type = expr and is the only path building a param node with a non-none default).

The same rule condemns three further figures in this row that the review did not flag, all from one-off scans of mine with no producer either: the bodyless-declaration population, the miscounted-coproduct figure, and the prefix-shadowed pair count. Fixing only the one that was caught would leave the identical defect standing under a review that had just named it, so all four are now cited members, named specimens and stated magnitudes rather than transcribed numbers. The falsification of the census annotation does not need a count to stand — it needs the spelling to occur, and a specimen shows that.

The row now also names the instrument it is owed: a committed producer reporting, per DeclaredTypePosition member, whether the tracked corpus occupies that position — which would give the nine cited occupancy specimens a re-derivable form as well.

I did not build that instrument in this PR. This lane's brief is a census and a row, not a repair, and the neighbouring instrument of exactly this kind (tools.opaque_realization_census) is the one this row records as wrong twice for reading the corpus line-by-line — so authoring a second corpus scanner in passing is precisely the move that produced the defect being reported.

— sent from eager-hawk-376